- Modify DmaPusher to use safe memory reads when handling compute
operations at High GPU accuracy
- Prevent potential memory corruption issues that could lead to
invalid dispatch parameters
- Previously, unsafe reads could result in corrupted launch_description
data in KeplerCompute::ProcessLaunch, causing invalid vkCmdDispatch
calls
- By enforcing safe reads specifically for compute operations, we
maintain performance for other GPU tasks while ensuring compute
dispatch stability
This change requires >= High GPU accuracy level to take effect.
This formats all copyright comments according to SPDX formatting guidelines.
Additionally, this resolves the remaining GPLv2 only licensed files by relicensing them to GPLv2.0-or-later.
Reimplement the buffer cache using cached bindings and page level
granularity for modification tracking. This also drops the usage of
shared pointers and virtual functions from the cache.
- Bindings are cached, allowing to skip work when the game changes few
bits between draws.
- OpenGL Assembly shaders no longer copy when a region has been modified
from the GPU to emulate constant buffers, instead GL_EXT_memory_object
is used to alias sub-buffers within the same allocation.
- OpenGL Assembly shaders stream constant buffer data using
glProgramBufferParametersIuivNV, from NV_parameter_buffer_object. In
theory this should save one hash table resolve inside the driver
compared to glBufferSubData.
- A new OpenGL stream buffer is implemented based on fences for drivers
that are not Nvidia's proprietary, due to their low performance on
partial glBufferSubData calls synchronized with 3D rendering (that
some games use a lot).
- Most optimizations are shared between APIs now, allowing Vulkan to
cache more bindings than before, skipping unnecesarry work.
This commit adds the necessary infrastructure to use Vulkan object from
OpenGL. Overall, it improves performance and fixes some bugs present on
the old cache. There are still some edge cases hit by some games that
harm performance on some vendors, this are planned to be fixed in later
commits.
This commit ensures that the host gpu is constantly fed with commands to
work with, while the guest gpu keeps producing the rest of the commands.
This reduces syncing time between host and guest gpu.
Instead of holding a reference that will get invalidated by
dma_pushbuffer.pop(), hold it as a copy. This doesn't have any
performance cost since CommandListHeader is 8 bytes long.
As fetching command list headers and and the list of command headers is a fixed 1:1 relation now, they can be implemented within a single call.
This cleans up the Step() logic quite a bit.
Fetching every u32 from memory leads to a big overhead. So let's fetch all of them as a block if possible.
This reduces the Memory::* calls by the dma_pusher by a factor of 10.
